Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Missing invoke handling causes infinite loop #1134

Open
wsmoses opened this issue Nov 8, 2023 · 14 comments
Open

Missing invoke handling causes infinite loop #1134

wsmoses opened this issue Nov 8, 2023 · 14 comments

Comments

@wsmoses
Copy link
Member

wsmoses commented Nov 8, 2023

See CI on 1.10 cc @vchuravy @gbaraldi
@wsmoses
Copy link
Member Author

wsmoses commented Nov 8, 2023

See https://github.com/EnzymeAD/Enzyme.jl/actions/runs/6786335843/job/18446651673

@wsmoses wsmoses mentioned this issue Nov 8, 2023
Closed
@vchuravy
Copy link
Member

vchuravy commented Nov 8, 2023 

WARNING: Method definition invsin(Any) in module Main at /home/runner/work/Enzyme.jl/Enzyme.jl/test/runtests.jl:515 overwritten at /home/runner/work/Enzyme.jl/Enzyme.jl/test/runtests.jl:1137.
generic: Error During Test at /home/runner/work/Enzyme.jl/Enzyme.jl/test/runtests.jl:1142
  Got exception outside of a @test
  StackOverflowError:
  Stacktrace:
       [1] runtime_generic_augfwd(::Type{Val{(false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true)}}, ::Val{1}, ::Val{(true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true)}, ::Val{@NamedTuple{1, 2, 3}}, ::typeof(hvcat), ::Nothing, ::NTuple{5, Int64}, ::Nothing, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64)
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:1433
       [2] hvcat
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/SparseArrays/src/sparsevector.jl:1269 [inlined]
       [3] hvcat
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/SparseArrays/src/sparsevector.jl:0 [inlined]
       [4] augmented_julia_hvcat_41331_inner_1wrap
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/SparseArrays/src/sparsevector.jl:0
       [5] macro expansion
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9836 [inlined]
       [6] enzyme_call
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9514 [inlined]
       [7] AugmentedForwardThunk
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9488 [inlined]
       [8] macro expansion
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:1390 [inlined]
  --- the last 8 lines are repeated 3876 more times ---
   [31017] runtime_generic_augfwd(::Type{Val{(false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true)}}, ::Val{1}, ::Val{(true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true)}, ::Val{@NamedTuple{1, 2, 3}}, ::typeof(hvcat), ::Nothing, ::NTuple{5, Int64}, ::Nothing, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64, ::Float64)
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:1433
   [31018] hvcat
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/SparseArrays/src/sparsevector.jl:1269 [inlined]
   [31019] hvcat
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/SparseArrays/src/sparsevector.jl:0 [inlined]
   [31020] augmented_julia_hvcat_41255_inner_1wrap
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/SparseArrays/src/sparsevector.jl:0
   [31021] macro expansion
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9836 [inlined]
   [31022] enzyme_call
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9514 [inlined]
   [31023] AugmentedForwardThunk
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9488 [inlined]
   [31024] macro expansion
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:1390 [inlined]
   [31025] runtime_generic_augfwd(::Type{Val{(false, false, true, false, false, false, false, false, false, true, false, false, false, false, false, false, true, false, false, false, false, false, false, true, false, false, false, false, false, false, true, false)}}, ::Val{1}, ::Val{(true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true)}, ::Val{@NamedTuple{1, 2, 3}}, ::typeof(hvcat), ::Nothing, ::NTuple{5, Int64}, ::Nothing, ::Float64, ::Float64, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Float64, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Float64, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Float64, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Nothing, ::Float64, ::Float64, ::Float64, ::Nothing)
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:1433
   [31026] hvcat
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/SparseArrays/src/sparsevector.jl:1269 [inlined]
   [31027] whocallsmorethan30args
         @ ~/work/Enzyme.jl/Enzyme.jl/test/runtests.jl:1170 [inlined]
   [31028] diffejulia_whocallsmorethan30args_41148wrap
         @ ~/work/Enzyme.jl/Enzyme.jl/test/runtests.jl:0
   [31029] macro expansion
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9836 [inlined]
   [31030] enzyme_call
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9514 [inlined]
   [31031] CombinedAdjointThunk
         @ Enzyme.Compiler ~/work/Enzyme.jl/Enzyme.jl/src/compiler.jl:9477 [inlined]
   [31032] autodiff(::ReverseMode{false, FFIABI}, f::Const{var"#whocallsmorethan30args#166"}, ::Type{Active}, args::Duplicated{Matrix{Float64}})
         @ Enzyme ~/work/Enzyme.jl/Enzyme.jl/src/Enzyme.jl:215
   [31033] autodiff(::ReverseMode{false, FFIABI}, ::var"#whocallsmorethan30args#166", ::Type, ::Duplicated{Matrix{Float64}})
         @ Enzyme ~/work/Enzyme.jl/Enzyme.jl/src/Enzyme.jl:224
   [31034] macro expansion
         @ ~/work/Enzyme.jl/Enzyme.jl/test/runtests.jl:1182 [inlined]
   [31035] macro expansion
         @ /opt/hostedtoolcache/julia/1.10-nightly/x64/share/julia/stdlib/v1.10/Test/src/Test.jl:1577 [inlined]
   [31036] top-level scope
         @ ~/work/Enzyme.jl/Enzyme.jl/test/runtests.jl:1143
   [31037] include(fname::String)
         @ Base.MainInclude ./client.jl:489
   [31038] top-level scope
         @ none:6
   [31039] eval
         @ Core ./boot.jl:383 [inlined]
   [31040] exec_options(opts::Base.JLOptions)
         @ Base ./client.jl:291
Test Summary: | Pass  Error  Total   Time
generic       |    8      1      9  11.6s

@vchuravy
Copy link
Member

vchuravy commented Nov 8, 2023 •
edited
Loading 

using Enzyme
using LinearAlgebra

function whocallsmorethan30args(R)
    temp = diag(R)
     R_inv = [temp[1] 0. 0. 0. 0. 0.;
         0. temp[2] 0. 0. 0. 0.;
         0. 0. temp[3] 0. 0. 0.;
         0. 0. 0. temp[4] 0. 0.;
         0. 0. 0. 0. temp[5] 0.;
     ]

    return sum(R_inv)
end

R = zeros(6,6)
dR = zeros(6, 6)
autodiff(Reverse, whocallsmorethan30args, Active, Duplicated(R, dR))

@vchuravy
Copy link
Member

vchuravy commented Nov 8, 2023

Hm can't reproduce locally. It was crashing on 1.10-beta, but is passing for me on rc1.

@vchuravy
Copy link
Member

vchuravy commented Nov 8, 2023

Even trying to directly reconstruct the call:

Enzyme.Compiler.runtime_generic_augfwd(
    Val{(false, false, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true)},
    Val{1}(),
    Val{(true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true, true)}(),
    Val{NamedTuple{(Symbol("1"), Symbol("2"), Symbol("3")),Tuple{Any,Any,Any}}}(),
    hvcat,
    nothing,
    (6,6,6,6,6), nothing,
    ntuple(_->1.0, 2*30)...)

@vchuravy vchuravy removed their assignment Nov 8, 2023
@vchuravy
Copy link
Member

vchuravy commented Nov 8, 2023

So after a short investigation. The code behaves properly outside the test-suite. It is not clear why this recurses to begin with.

@wsmoses
Copy link
Member Author

wsmoses commented Nov 29, 2023

This reproduces in a file for me:

using Enzyme
using Statistics

Enzyme.API.printall!(true)

    function whocallsmorethan30args(temp)
         R_inv = [temp 0. 0. 0. 0. 0.; 
             0. temp 0. 0. 0. 0.; 
             0. 0. temp 0. 0. 0.; 
             0. 0. 0. temp 0. 0.; 
             0. 0. 0. 0. temp 0.; 
         ]
    
	 return R_inv[1,1]
    end
    
    autodiff(Forward, whocallsmorethan30args, DuplicatedNoNeed, Duplicated(1.0,1.0))

I think I know what is happening and why, but not immediately a good fix.

Right now applylatest/applygeneric/invoke all become a applygeneric on runtime_generic_augfwd. However, in 1.10 the first call above becomes a jl_invoke(hvcat, ....) to a specialized method for all args. That jl_invoke becomes a call to autodiff hvcat(all those args). Once again the julia compiler decides to wrap that inside of a jl_invoke, instead of doing the specialized one, creating the loop.

@wsmoses
Copy link
Member Author

wsmoses commented Nov 29, 2023

In my opinion the best solution is for us to pass the mi down to thunk, but we cannot since a generated method cannot take a methodinstance in a val [which I know concerns you for other reasons, but thats my immediate thoughts fo what blocks]

@wsmoses
Copy link
Member Author

wsmoses commented Nov 29, 2023

This PR does a partial fix but hits a julia issue I don't understand with gpucompiler.jl: #1169

Specifically the outermost function is this, which has a vararg at the end and has a different julia calling abi:

svec(typeof(hvcat), NTuple{5, Int64}, Float64, Float64, Vararg{Float64})

Could gpucompiler give us one with the usual ABI?

define nonnull {} addrspace(10)* @japi1_hvcat_3138({} addrspace(10)* %function, {} addrspace(10)** noalias nocapture noundef readonly %args, i32 %nargs) #0 !dbg !4 {
top:
  %stackargs = alloca {} addrspace(10)**, align 8
  store volatile {} addrspace(10)** %args, {} addrspace(10)*** %stackargs, align 8
  %0 = call {}*** @julia.get_pgcstack()
  %1 = bitcast {}*** %0 to {}**
  %current_task = getelementptr inbounds {}*, {}** %1, i64 -14
  %2 = bitcast {}** %current_task to i64*
  %world_age = getelementptr inbounds i64, i64* %2, i64 15
  %3 = getelementptr inbounds {} addrspace(10)*, {} addrspace(10)** %args, i32 0
  %4 = load {} addrspace(10)*, {} addrspace(10)** %3, align 8, !tbaa !8, !invariant.load !7, !alias.scope !12, !noalias !15, !nonnull !7, !dereferenceable !20, !align !21
  %5 = getelementptr inbounds {} addrspace(10)*, {} addrspace(10)** %args, i64 1
  %6 = sub i32 %nargs, 1
  %7 = bitcast {}*** %0 to {}**
  %current_task1 = getelementptr inbounds {}*, {}** %7, i64 -14
  %ptls_field = getelementptr inbounds {}*, {}** %current_task1, i64 16
  %ptls_load = load {}*, {}** %ptls_field, align 8, !tbaa !22
  %ptls = bitcast {}* %ptls_load to {}**
  %8 = bitcast {}** %ptls to i64**
  %9 = getelementptr inbounds i64*, i64** %8, i64 2
  %safepoint = load i64*, i64** %9, align 8, !tbaa !8, !invariant.load !7
  fence syncscope("singlethread") seq_cst
  call void @julia.safepoint(i64* %safepoint), !dbg !24
  fence syncscope("singlethread") seq_cst
  %10 = addrspacecast {} addrspace(10)* %4 to {} addrspace(11)*, !dbg !25
  %11 = bitcast {} addrspace(11)* %10 to [5 x i64] addrspace(11)*, !dbg !25
  %12 = getelementptr inbounds [5 x i64], [5 x i64] addrspace(11)* %11, i32 0, i32 0, !dbg !25
  %unbox = load i64, i64 addrspace(11)* %12, align 8, !dbg !29, !tbaa !33, !alias.scope !37, !noalias !38
  %13 = call nonnull {} addrspace(10)* @ijl_alloc_array_2d({} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515882277456 to {}*) to {} addrspace(10)*), i64 5, i64 %unbox), !dbg !29
  %14 = addrspacecast {} addrspace(10)* %13 to {} addrspace(11)*, !dbg !39
  %15 = bitcast {} addrspace(11)* %14 to { i8 addrspace(13)*, i64, i16, i16, i32 } addrspace(11)*, !dbg !39
  %arraylen_ptr = getelementptr inbounds { i8 addrspace(13)*, i64, i16, i16, i32 }, { i8 addrspace(13)*, i64, i16, i16, i32 } addrspace(11)* %15, i32 0, i32 1, !dbg !39
  %arraylen = load i64, i64 addrspace(11)* %arraylen_ptr, align 8, !dbg !39, !tbaa !8, !range !43, !invariant.load !7, !alias.scope !12, !noalias !15
  %16 = sub i32 %nargs, 1, !dbg !44
  %17 = sext i32 %16 to i64, !dbg !44
  %18 = icmp eq i64 %arraylen, %17, !dbg !46
  %19 = zext i1 %18 to i8, !dbg !46
  %20 = trunc i8 %19 to i1, !dbg !52
  %21 = xor i1 %20, true, !dbg !52
  %22 = zext i1 %21 to i8, !dbg !42
  %23 = trunc i8 %22 to i1, !dbg !42
  %24 = xor i1 %23, true, !dbg !42
  br i1 %24, label %L14, label %L8, !dbg !42

L8:                                               ; preds = %top
  %25 = addrspacecast {} addrspace(10)* %13 to {} addrspace(11)*, !dbg !55
  %26 = bitcast {} addrspace(11)* %25 to { i8 addrspace(13)*, i64, i16, i16, i32 } addrspace(11)*, !dbg !55
  %arraylen_ptr2 = getelementptr inbounds { i8 addrspace(13)*, i64, i16, i16, i32 }, { i8 addrspace(13)*, i64, i16, i16, i32 } addrspace(11)* %26, i32 0, i32 1, !dbg !55
  %arraylen3 = load i64, i64 addrspace(11)* %arraylen_ptr2, align 8, !dbg !55, !tbaa !8, !range !43, !invariant.load !7, !alias.scope !12, !noalias !15
  %27 = sub i32 %nargs, 1, !dbg !57
  %28 = sext i32 %27 to i64, !dbg !57
  %29 = call nonnull {} addrspace(10)* @ijl_box_int64(i64 signext %arraylen3), !dbg !56
  %30 = call nonnull {} addrspace(10)* @ijl_box_int64(i64 signext %28), !dbg !56
  %31 = call nonnull {} addrspace(10)* ({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32, {} addrspace(10)*)*, {} addrspace(10)*, {} addrspace(10)*, ...) @julia.call2({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32, {} addrspace(10)*)* @ijl_invoke, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515739994608 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515715232560 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515820605968 to {}*) to {} addrspace(10)*), {} addrspace(10)* %29, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515820605936 to {}*) to {} addrspace(10)*), {} addrspace(10)* %30, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515820605904 to {}*) to {} addrspace(10)*)) #6, !dbg !56
  %32 = insertvalue [1 x {} addrspace(10)*] zeroinitializer, {} addrspace(10)* %31, 0, !dbg !58
  %33 = bitcast {}*** %0 to {}**, !dbg !56
  %current_task4 = getelementptr inbounds {}*, {}** %33, i64 -14, !dbg !56
  %box = call noalias nonnull dereferenceable(8) {} addrspace(10)* @julia.gc_alloc_obj({}** %current_task4, i64 8, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515785584480 to {}*) to {} addrspace(10)*)) #7, !dbg !56
  %34 = bitcast {} addrspace(10)* %box to [1 x {} addrspace(10)*] addrspace(10)*, !dbg !56
  store [1 x {} addrspace(10)*] %32, [1 x {} addrspace(10)*] addrspace(10)* %34, align 8, !dbg !56, !tbaa !33, !alias.scope !37, !noalias !38
  %35 = addrspacecast {} addrspace(10)* %box to {} addrspace(12)*, !dbg !56
  call void @ijl_throw({} addrspace(12)* %35), !dbg !56
  unreachable, !dbg !56

L14:                                              ; preds = %top
  br i1 false, label %L72, label %L14.L15_crit_edge, !dbg !60

L14.L15_crit_edge:                                ; preds = %L14
  br label %L15, !dbg !60

L15:                                              ; preds = %L14.L15_crit_edge, %L71
  %value_phi = phi i64 [ 1, %L14.L15_crit_edge ], [ %value_phi27, %L71 ]
  %value_phi5 = phi i64 [ 1, %L14.L15_crit_edge ], [ %value_phi28, %L71 ]
  %value_phi6 = phi i64 [ 1, %L14.L15_crit_edge ], [ %value_phi26, %L71 ]
  %36 = sub i64 %value_phi, 1, !dbg !61
  %37 = addrspacecast {} addrspace(10)* %4 to {} addrspace(11)*, !dbg !61
  %38 = bitcast {} addrspace(11)* %37 to i64 addrspace(11)*, !dbg !61
  %39 = getelementptr inbounds i64, i64 addrspace(11)* %38, i64 %36, !dbg !61
  %unbox7 = load i64, i64 addrspace(11)* %12, align 8, !dbg !63, !tbaa !33, !alias.scope !37, !noalias !38
  %unbox8 = load i64, i64 addrspace(11)* %39, align 8, !dbg !63, !tbaa !33, !alias.scope !37, !noalias !38
  %40 = icmp eq i64 %unbox7, %unbox8, !dbg !63
  %41 = zext i1 %40 to i8, !dbg !63
  %42 = trunc i8 %41 to i1, !dbg !65
  %43 = xor i1 %42, true, !dbg !65
  %44 = zext i1 %43 to i8, !dbg !62
  %45 = trunc i8 %44 to i1, !dbg !62
  %46 = xor i1 %45, true, !dbg !62
  br i1 %46, label %L27, label %L22, !dbg !62

L22:                                              ; preds = %L15
  %47 = sub i64 %value_phi, 1, !dbg !66
  %48 = addrspacecast {} addrspace(10)* %4 to {} addrspace(11)*, !dbg !66
  %49 = bitcast {} addrspace(11)* %48 to i64 addrspace(11)*, !dbg !66
  %50 = getelementptr inbounds i64, i64 addrspace(11)* %49, i64 %47, !dbg !66
  %51 = call nonnull {} addrspace(10)* @ijl_box_int64(i64 signext %value_phi), !dbg !67
  %unbox9 = load i64, i64 addrspace(11)* %12, align 8, !dbg !67, !tbaa !33, !alias.scope !37, !noalias !38
  %52 = call nonnull {} addrspace(10)* @ijl_box_int64(i64 signext %unbox9), !dbg !67
  %unbox10 = load i64, i64 addrspace(11)* %50, align 8, !dbg !67, !tbaa !33, !alias.scope !37, !noalias !38
  %53 = call nonnull {} addrspace(10)* @ijl_box_int64(i64 signext %unbox10), !dbg !67
  %54 = call nonnull {} addrspace(10)* ({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32, {} addrspace(10)*)*, {} addrspace(10)*, {} addrspace(10)*, ...) @julia.call2({} addrspace(10)* ({} addrspace(10)*, {} addrspace(10)**, i32, {} addrspace(10)*)* @ijl_invoke, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515739994608 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515715232560 to {}*) to {} addrspace(10)*), {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515820605872 to {}*) to {} addrspace(10)*), {} addrspace(10)* %51, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515820605808 to {}*) to {} addrspace(10)*), {} addrspace(10)* %52, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515820605936 to {}*) to {} addrspace(10)*), {} addrspace(10)* %53, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515820605904 to {}*) to {} addrspace(10)*)) #6, !dbg !67
  %55 = insertvalue [1 x {} addrspace(10)*] zeroinitializer, {} addrspace(10)* %54, 0, !dbg !68
  %56 = bitcast {}*** %0 to {}**, !dbg !67
  %current_task11 = getelementptr inbounds {}*, {}** %56, i64 -14, !dbg !67
  %box12 = call noalias nonnull dereferenceable(8) {} addrspace(10)* @julia.gc_alloc_obj({}** %current_task11, i64 8, {} addrspace(10)* addrspacecast ({}* inttoptr (i64 140515761616736 to {}*) to {} addrspace(10)*)) #7, !dbg !67
  %57 = bitcast {} addrspace(10)* %box12 to [1 x {} addrspace(10)*] addrspace(10)*, !dbg !67
  store [1 x {} addrspace(10)*] %55, [1 x {} addrspace(10)*] addrspace(10)* %57, align 8, !dbg !67, !tbaa !33, !alias.scope !37, !noalias !38
  %58 = addrspacecast {} addrspace(10)* %box12 to {} addrspace(12)*, !dbg !67
  call void @ijl_throw({} addrspace(12)* %58), !dbg !67
  unreachable, !dbg !67

L27:                                              ; preds = %L15
  %unbox15 = load i64, i64 addrspace(11)* %12, align 8, !dbg !71, !tbaa !33, !alias.scope !37, !noalias !38
  %59 = icmp sle i64 1, %unbox15, !dbg !71
  %60 = zext i1 %59 to i8, !dbg !76
  %61 = trunc i8 %60 to i1, !dbg !76
  %62 = xor i1 %61, true, !dbg !76
  br i1 %62, label %L30, label %L29, !dbg !76

L29:                                              ; preds = %L27
  %63 = icmp ne i64 addrspace(11)* %12, null
  br i1 %63, label %guard_pass, label %guard_exit

L30:                                              ; preds = %L27
  br label %L31, !dbg !76

L31:                                              ; preds = %L30, %guard_exit31
  %value_phi16 = phi i64 [ %guard_res33, %guard_exit31 ], [ 0, %L30 ]
  br label %L33, !dbg !79

L33:                                              ; preds = %L31
  br label %L34, !dbg !81

L34:                                              ; preds = %L33
  %64 = icmp slt i64 %value_phi16, 1, !dbg !84
  %65 = zext i1 %64 to i8, !dbg !90
  %66 = trunc i8 %65 to i1, !dbg !90
  %67 = xor i1 %66, true, !dbg !90
  br i1 %67, label %L37, label %L36, !dbg !90

L36:                                              ; preds = %L34
  br label %L38, !dbg !90

L37:                                              ; preds = %L34
  br label %L38, !dbg !90

L38:                                              ; preds = %L37, %L36
  %value_phi17 = phi i8 [ 1, %L36 ], [ 0, %L37 ]
  %value_phi18 = phi i64 [ 1, %L37 ], [ undef, %L36 ]
  %value_phi19 = phi i64 [ 1, %L37 ], [ undef, %L36 ]
  %68 = trunc i8 %value_phi17 to i1, !dbg !83
  %69 = xor i1 %68, true, !dbg !83
  %70 = zext i1 %69 to i8, !dbg !83
  %71 = trunc i8 %70 to i1, !dbg !83
  %72 = xor i1 %71, true, !dbg !83
  br i1 %72, label %L38.L60_crit_edge, label %L38.L43_crit_edge, !dbg !83

L38.L60_crit_edge:                                ; preds = %L38
  br label %L60, !dbg !92

L38.L43_crit_edge:                                ; preds = %L38
  br label %L43, !dbg !90

L43:                                              ; preds = %L38.L43_crit_edge, %L59
  %value_phi20 = phi i64 [ %value_phi6, %L38.L43_crit_edge ], [ %93, %L59 ]
  %value_phi21 = phi i64 [ %value_phi18, %L38.L43_crit_edge ], [ %value_phi23, %L59 ]
  %value_phi22 = phi i64 [ %value_phi19, %L38.L43_crit_edge ], [ %value_phi24, %L59 ]
  %73 = sub i32 %nargs, 1, !dbg !94
  %74 = sext i32 %73 to i64, !dbg !94
  %75 = getelementptr inbounds {} addrspace(10)*, {} addrspace(10)** %args, i64 1, !dbg !94
  %76 = sub i64 %value_phi20, 1, !dbg !94
  %77 = add i64 %76, 1, !dbg !94
  %78 = getelementptr inbounds {} addrspace(10)*, {} addrspace(10)** %args, i64 %77, !dbg !94
  %getfield = load {} addrspace(10)*, {} addrspace(10)** %78, align 8, !dbg !94, !tbaa !96, !alias.scope !37, !noalias !38, !nonnull !7, !dereferenceable !21, !align !21
  %79 = sub i64 %value_phi, 1, !dbg !97
  %80 = mul i64 %79, 1, !dbg !97
  %81 = add i64 0, %80, !dbg !97
  %82 = addrspacecast {} addrspace(10)* %13 to {} addrspace(11)*, !dbg !97
  %83 = bitcast {} addrspace(11)* %82 to {} addrspace(10)* addrspace(11)*, !dbg !97
  %arraysize_ptr = getelementptr inbounds {} addrspace(10)*, {} addrspace(10)* addrspace(11)* %83, i32 3, !dbg !97
  %84 = bitcast {} addrspace(10)* addrspace(11)* %arraysize_ptr to i64 addrspace(11)*, !dbg !97
  %arraysize = load i64, i64 addrspace(11)* %84, align 8, !dbg !97, !tbaa !8, !range !43, !invariant.load !7, !alias.scope !12, !noalias !15
  %stride = mul i64 1, %arraysize, !dbg !97
  %85 = sub i64 %value_phi21, 1, !dbg !97
  %86 = mul i64 %85, %stride, !dbg !97
  %87 = add i64 %81, %86, !dbg !97
  %88 = addrspacecast {} addrspace(10)* %13 to {} addrspace(11)*, !dbg !97
  %89 = bitcast {} addrspace(11)* %88 to { i8 addrspace(13)*, i64, i16, i16, i32 } addrspace(11)*, !dbg !97
  %arrayptr_ptr = getelementptr inbounds { i8 addrspace(13)*, i64, i16, i16, i32 }, { i8 addrspace(13)*, i64, i16, i16, i32 } addrspace(11)* %89, i32 0, i32 0, !dbg !97
  %arrayptr = load i8 addrspace(13)*, i8 addrspace(13)* addrspace(11)* %arrayptr_ptr, align 8, !dbg !97, !tbaa !8, !invariant.load !7, !alias.scope !12, !noalias !15, !nonnull !7
  %90 = bitcast i8 addrspace(13)* %arrayptr to double addrspace(13)*, !dbg !97
  %memcpy_refined_dst = getelementptr inbounds double, double addrspace(13)* %90, i64 %87, !dbg !97
  %91 = addrspacecast {} addrspace(10)* %getfield to {} addrspace(11)*, !dbg !97
  %memcpy_refined_src = bitcast {} addrspace(11)* %91 to double addrspace(11)*, !dbg !97
  %92 = load double, double addrspace(11)* %memcpy_refined_src, align 8, !dbg !97, !tbaa !33, !alias.scope !37, !noalias !38
  store double %92, double addrspace(13)* %memcpy_refined_dst, align 8, !dbg !97, !tbaa !99, !alias.scope !37, !noalias !38
  %93 = add i64 %value_phi20, 1, !dbg !101
  %94 = icmp eq i64 %value_phi22, %value_phi16, !dbg !104
  %95 = zext i1 %94 to i8, !dbg !104
  %96 = trunc i8 %95 to i1, !dbg !92
  %97 = xor i1 %96, true, !dbg !92
  br i1 %97, label %L52, label %L51, !dbg !92

L51:                                              ; preds = %L43
  br label %L54, !dbg !92

L52:                                              ; preds = %L43
  %98 = add i64 %value_phi22, 1, !dbg !105
  br label %L54, !dbg !92

L54:                                              ; preds = %L52, %L51
  %value_phi23 = phi i64 [ %98, %L52 ], [ undef, %L51 ]
  %value_phi24 = phi i64 [ %98, %L52 ], [ undef, %L51 ]
  %value_phi25 = phi i8 [ 1, %L51 ], [ 0, %L52 ]
  %99 = trunc i8 %value_phi25 to i1, !dbg !93
  %100 = xor i1 %99, true, !dbg !93
  %101 = zext i1 %100 to i8, !dbg !93
  %102 = trunc i8 %101 to i1, !dbg !93
  %103 = xor i1 %102, true, !dbg !93
  br i1 %103, label %L54.L60_crit_edge, label %L59, !dbg !93

L54.L60_crit_edge:                                ; preds = %L54
  br label %L60, !dbg !92

L59:                                              ; preds = %L54
  br label %L43, !dbg !90

L60:                                              ; preds = %L38.L60_crit_edge, %L54.L60_crit_edge
  %value_phi26 = phi i64 [ %93, %L54.L60_crit_edge ], [ %value_phi6, %L38.L60_crit_edge ]
  %104 = icmp eq i64 %value_phi5, 5, !dbg !107
  %105 = zext i1 %104 to i8, !dbg !107
  %106 = trunc i8 %105 to i1, !dbg !108
  %107 = xor i1 %106, true, !dbg !108
  br i1 %107, label %L64, label %L63, !dbg !108

L63:                                              ; preds = %L60
  br label %L66, !dbg !108

L64:                                              ; preds = %L60
  %108 = add i64 %value_phi5, 1, !dbg !110
  br label %L66, !dbg !108

L66:                                              ; preds = %L64, %L63
  %value_phi27 = phi i64 [ %108, %L64 ], [ undef, %L63 ]
  %value_phi28 = phi i64 [ %108, %L64 ], [ undef, %L63 ]
  %value_phi29 = phi i8 [ 1, %L63 ], [ 0, %L64 ]
  %109 = trunc i8 %value_phi29 to i1, !dbg !109
  %110 = xor i1 %109, true, !dbg !109
  %111 = zext i1 %110 to i8, !dbg !109
  %112 = trunc i8 %111 to i1, !dbg !109
  %113 = xor i1 %112, true, !dbg !109
  br i1 %113, label %L72, label %L71, !dbg !109

L71:                                              ; preds = %L66
  br label %L15, !dbg !109

L72:                                              ; preds = %L66, %L14
  ret {} addrspace(10)* %13, !dbg !112

after_throw:                                      ; No predecessors!
  call void @llvm.trap(), !dbg !56
  unreachable, !dbg !56

after_noret:                                      ; No predecessors!
  call void @llvm.trap(), !dbg !56
  unreachable, !dbg !56

after_throw13:                                    ; No predecessors!
  call void @llvm.trap(), !dbg !67
  unreachable, !dbg !67

after_noret14:                                    ; No predecessors!
  call void @llvm.trap(), !dbg !67
  unreachable, !dbg !67

guard_pass:                                       ; preds = %L29
  br label %guard_exit

guard_exit:                                       ; preds = %guard_pass, %L29
  %guard_res = phi i1 [ false, %L29 ], [ true, %guard_pass ]
  br i1 %guard_res, label %guard_pass30, label %guard_exit31

guard_pass30:                                     ; preds = %guard_exit
  %unbox32 = load i64, i64 addrspace(11)* %12, align 8, !tbaa !33, !alias.scope !37, !noalias !38
  br label %guard_exit31

guard_exit31:                                     ; preds = %guard_pass30, %guard_exit
  %guard_res33 = phi i64 [ undef, %guard_exit ], [ %unbox32, %guard_pass30 ]
  br label %L31, !dbg !76
}

@wsmoses
Copy link
Member Author

wsmoses commented Nov 29, 2023

@vchuravy it looks like in get_function_name of codegen.cpp, the specsig flag, which should be set to true by gpucompiler is getting ignored or overridden or something

@wsmoses
Copy link
Member Author

wsmoses commented Nov 29, 2023

@gbaraldi @vtjnash if you have any insights here, especially on why 1.10-rc1 broke and caused this infinite recursion to begin with [whereas beta3 was fine]

1 similar comment
@wsmoses
Copy link
Member Author

wsmoses commented Nov 29, 2023

@gbaraldi @vtjnash if you have any insights here, especially on why 1.10-rc1 broke and caused this infinite recursion to begin with [whereas beta3 was fine]

@vchuravy
Copy link
Member

vchuravy commented Dec 4, 2023 •
edited
Loading

t looks like in get_function_name of codegen.cpp, the specsig flag, which should be set to true by gpucompiler is getting ignored or overridden or something

Can you expand on this? Who is the caller? get_function_name is called from emit_function.

Is params.params->prefer_specsig true or false? At this location https://github.com/JuliaLang/julia/blob/dbb9c46795b0e45b770254542854526f3db119b5/src/codegen.cpp#L7273

@vchuravy
Copy link
Member

So the culprit is that we are not handling the invoke correctly that was added in JuliaSparse/SparseArrays.jl@c402d09

hvcat(rows::Tuple{Vararg{Int}}, n1::Number, ns::Vararg{Number}) = invoke(hvcat, Tuple{typeof(rows), Vararg{Number}}, rows, n1, ns...)
hvcat(rows::Tuple{Vararg{Int}}, n1::N, ns::Vararg{N}) where {N<:Number} = invoke(hvcat, Tuple{typeof(rows), Vararg{N}}, rows, n1, ns...)

which got backported to 1.10-rc1 JuliaLang/julia#48977

@vchuravy vchuravy changed the title Julia 1.10 infinite recursion Missing invoke handling causes infinite loop Dec 15, 2023
@vchuravy vchuravy mentioned this issue Dec 15, 2023
Open
@wsmoses wsmoses mentioned this issue Feb 8, 2024
Open
@wsmoses wsmoses mentioned this issue Mar 6, 2024
Closed
@wsmoses wsmoses mentioned this issue Apr 17, 2024
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@vchuravy @wsmoses